{
  "id": 313338,
  "title": "Starter pack with tez",
  "url": "/competitions/ultra-mnist/discussion/313338",
  "author_name": "Konrad Banachewicz",
  "post_date": "2022-03-16T15:08:13.787000",
  "votes": 16,
  "comment_count": 3,
  "views": 0,
  "content": "<p>After initial turbulence I threw away my TF code and started from scratch using tez - a rather neat PyTorch trainer written by <a href=\"https://www.kaggle.com/abhishek\" target=\"_blank\">@abhishek</a> (because who else). If you feel like giving it a shot, here are some resources:</p>\n<p><strong>Notebooks</strong></p>\n<p>training: <a href=\"https://www.kaggle.com/konradb/tez-model-train\" target=\"_blank\">https://www.kaggle.com/konradb/tez-model-train</a><br>\ninference: <a href=\"https://www.kaggle.com/konradb/tez-model-inference\" target=\"_blank\">https://www.kaggle.com/konradb/tez-model-inference</a></p>\n<p><strong>Resized datasets</strong></p>\n<p>128x128: <a href=\"https://www.kaggle.com/konradb/ultramnist-resized-128\" target=\"_blank\">https://www.kaggle.com/konradb/ultramnist-resized-128</a><br>\n256x256: <a href=\"https://www.kaggle.com/konradb/ultramnist-resized-256\" target=\"_blank\">https://www.kaggle.com/konradb/ultramnist-resized-256</a><br>\n512x512: <a href=\"https://www.kaggle.com/konradb/ultramnist-resized-512\" target=\"_blank\">https://www.kaggle.com/konradb/ultramnist-resized-512</a></p>\n<p><strong>Fitted models</strong></p>\n<p><a href=\"https://www.kaggle.com/konradb/tez-resnet50-folds\" target=\"_blank\">https://www.kaggle.com/konradb/tez-resnet50-folds</a></p>\n<p>The validation score seems rather nice, but LB is lower. Still hunting that bug.</p>",
  "messages": [
    {
      "id": 1724846,
      "postDate": "2022-03-16T15:08:13.787Z",
      "content": "<p>After initial turbulence I threw away my TF code and started from scratch using tez - a rather neat PyTorch trainer written by <a href=\"https://www.kaggle.com/abhishek\" target=\"_blank\">@abhishek</a> (because who else). If you feel like giving it a shot, here are some resources:</p>\n<p><strong>Notebooks</strong></p>\n<p>training: <a href=\"https://www.kaggle.com/konradb/tez-model-train\" target=\"_blank\">https://www.kaggle.com/konradb/tez-model-train</a><br>\ninference: <a href=\"https://www.kaggle.com/konradb/tez-model-inference\" target=\"_blank\">https://www.kaggle.com/konradb/tez-model-inference</a></p>\n<p><strong>Resized datasets</strong></p>\n<p>128x128: <a href=\"https://www.kaggle.com/konradb/ultramnist-resized-128\" target=\"_blank\">https://www.kaggle.com/konradb/ultramnist-resized-128</a><br>\n256x256: <a href=\"https://www.kaggle.com/konradb/ultramnist-resized-256\" target=\"_blank\">https://www.kaggle.com/konradb/ultramnist-resized-256</a><br>\n512x512: <a href=\"https://www.kaggle.com/konradb/ultramnist-resized-512\" target=\"_blank\">https://www.kaggle.com/konradb/ultramnist-resized-512</a></p>\n<p><strong>Fitted models</strong></p>\n<p><a href=\"https://www.kaggle.com/konradb/tez-resnet50-folds\" target=\"_blank\">https://www.kaggle.com/konradb/tez-resnet50-folds</a></p>\n<p>The validation score seems rather nice, but LB is lower. Still hunting that bug.</p>",
      "rawMarkdown": "After initial turbulence I threw away my TF code and started from scratch using tez - a rather neat PyTorch trainer written by @abhishek (because who else). If you feel like giving it a shot, here are some resources:\n\n**Notebooks**\n\ntraining: https://www.kaggle.com/konradb/tez-model-train\ninference: https://www.kaggle.com/konradb/tez-model-inference\n\n**Resized datasets**\n\n128x128: https://www.kaggle.com/konradb/ultramnist-resized-128\n256x256: https://www.kaggle.com/konradb/ultramnist-resized-256\n512x512: https://www.kaggle.com/konradb/ultramnist-resized-512\n\n**Fitted models**\n\nhttps://www.kaggle.com/konradb/tez-resnet50-folds\n\nThe validation score seems rather nice, but LB is lower. Still hunting that bug.",
      "votes": 16
    },
    {
      "id": 1725164,
      "postDate": "2022-03-16T21:35:18.240Z",
      "content": "<p>Been there done that on training. </p>\n<p>Kagglers should support and RESPECT experienced professionals that are still contributing for the growing of this Platform.</p>\n<p>Users should justify their downvotes. </p>",
      "rawMarkdown": "Been there done that on training. \n\nKagglers should support and RESPECT experienced professionals that are still contributing for the growing of this Platform.\n\nUsers should justify their downvotes. ",
      "votes": 2
    },
    {
      "id": 1725018,
      "postDate": "2022-03-16T18:26:49.960Z",
      "rawMarkdown": "",
      "isDeleted": true,
      "replies": [
        {
          "id": 1725024,
          "postDate": "2022-03-16T18:36:49.177Z",
          "content": "<p>i always upvote good work and especially when someone uses Tez hehe… i should write: upvotes are not endorsements ;)</p>",
          "rawMarkdown": "i always upvote good work and especially when someone uses Tez hehe... i should write: upvotes are not endorsements ;)",
          "votes": 2
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1725164,
      "author_name": "Marília Prata",
      "author_url": "",
      "post_date": "2022-03-16T21:35:18.240000",
      "content": "<p>Been there done that on training. </p>\n<p>Kagglers should support and RESPECT experienced professionals that are still contributing for the growing of this Platform.</p>\n<p>Users should justify their downvotes. </p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1725018,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-03-16T18:26:49.960000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 1725024,
          "author_name": "Abhishek Thakur",
          "author_url": "",
          "post_date": "2022-03-16T18:36:49.177000",
          "content": "<p>i always upvote good work and especially when someone uses Tez hehe… i should write: upvotes are not endorsements ;)</p>",
          "votes": 2,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1724846": "After initial turbulence I threw away my TF code and started from scratch using tez - a rather neat PyTorch trainer written by @abhishek (because who else). If you feel like giving it a shot, here are some resources:\n\n**Notebooks**\n\ntraining: https://www.kaggle.com/konradb/tez-model-train\ninference: https://www.kaggle.com/konradb/tez-model-inference\n\n**Resized datasets**\n\n128x128: https://www.kaggle.com/konradb/ultramnist-resized-128\n256x256: https://www.kaggle.com/konradb/ultramnist-resized-256\n512x512: https://www.kaggle.com/konradb/ultramnist-resized-512\n\n**Fitted models**\n\nhttps://www.kaggle.com/konradb/tez-resnet50-folds\n\nThe validation score seems rather nice, but LB is lower. Still hunting that bug.",
    "1725164": "Been there done that on training. \n\nKagglers should support and RESPECT experienced professionals that are still contributing for the growing of this Platform.\n\nUsers should justify their downvotes. ",
    "1725018": ""
  }
}